Centralize CHUNK and D constants as single source of truth

Problem

CHUNK=16 and D=128 appear as magic numbers in multiple locations:

• csrc/flash_kda.cpp — static_assert and workspace size calculation
• csrc/smxx/fwd_kernel1.cuh — kernel constants
• csrc/smxx/fwd_kernel2.cuh — kernel constants
• tests/torch_ref.py — reference implementation
• tests/test_fwd.py — test parameters
• docs/20260420-flashkda-v1-deep-dive.md — documentation

There is no single source of truth. If someone wanted to support D=64 or D=256, they'd need to change dozens of locations.

Recommendation

Define CHUNK and D as compile-time constants in one place (e.g., setup.py as compiler defines, or a shared header), and reference them everywhere else. Add a test that verifies the error message when D != 128.

Impact

Extensibility. Makes it possible to support different D values without a multi-file hunt for magic numbers.

Start by reading the constant definitions and uses in csrc/flash_kda.cpp, csrc/smxx/fwd_kernel1.cuh, csrc/smxx/fwd_kernel2.cuh, tests/torch_ref.py, tests/test_fwd.py, and docs/20260420-flashkda-v1-deep-dive.md. Choose a shared compile-time source, update the listed references, and add a test covering the D != 128 error message. Done means the constants are defined once and the affected tests pass.